Robust Perceptual Image Hashing for Screen-Shooting Attack

In this paper, we propose a Swin Transformer-based perceptual image hashing scheme to resist the impact of screen-shooting distortion on content authentication. The system applies a Swin Transformer-based U-shaped architecture to form a feature extractor, cooperates through encoder-decoder for feature extraction and compression, and employs a hash generator consisting of several fully connected layers to generate the final hash sequence. In addition, the scheme adds a distortion simulator to the training process, which uses image processing operations to simulate distortion on the original image to assist the feature extractor and hash generator in learning to resist screen-shooting distortion. Our experiments show that the proposed hashing is robust to single and hybrid content-preserving operations in screen-shooting attack. In addition, we also demonstrate that our hashing scheme outperforms some state-of-the-art schemes in terms of robustness and discrimination.